ab_off_r0.log strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 1.54 conv 0.18 ab 0.36 z 1.00 rec 0.49 out-proj 0.96 hc-read1+router 2.80 shared+quant 1.30 waitA 1.90 VRAM hits 3.18 waitB 0.45 PCIe grp 0.31 waitCPU 6.28 copy+combine 0.49 (gap) 0.05 head 0.95   hc0 norm 0.47   hc0 down 0.77   hc0 up 0.78 | QSA layers: q8+kv-idx 0.11 k/v+norm-rope 0.23 kv+idx append 0.16 q+q-idx 0.73 scores+topk 0.56 kv-resolve 0.34 attention 0.60 gate 0.05 out-proj 0.36 hc-read1+router 0.96 shared+quant 0.44 waitA 0.15 VRAM hits 1.04 waitB 0.19 PCIe grp 0.11 waitCPU 2.90 copy+combine 0.17 (gap) 0.02   hc0 norm 0.10   hc0 down 0.26   hc0 up 0.26 | total 34.01 ms/window over 62 windows
ab_off_r1.log strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 1.59 conv 0.18 ab 0.36 z 1.03 rec 0.49 out-proj 0.98 hc-read1+router 2.88 shared+quant 1.32 waitA 1.74 VRAM hits 3.39 waitB 0.73 PCIe grp 0.32 waitCPU 6.99 copy+combine 0.50 (gap) 0.05 head 0.98   hc0 norm 0.48   hc0 down 0.80   hc0 up 0.81 | QSA layers: q8+kv-idx 0.11 k/v+norm-rope 0.23 kv+idx append 0.16 q+q-idx 0.76 scores+topk 0.57 kv-resolve 0.36 attention 0.62 gate 0.05 out-proj 0.37 hc-read1+router 0.99 shared+quant 0.45 waitA 0.18 VRAM hits 1.10 waitB 0.32 PCIe grp 0.11 waitCPU 3.31 copy+combine 0.18 (gap) 0.02   hc0 norm 0.10   hc0 down 0.27   hc0 up 0.27 | total 36.17 ms/window over 68 windows
ab_off_r2.log strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 1.57 conv 0.18 ab 0.36 z 1.02 rec 0.50 out-proj 0.97 hc-read1+router 2.85 shared+quant 1.30 waitA 1.63 VRAM hits 3.35 waitB 0.65 PCIe grp 0.32 waitCPU 6.18 copy+combine 0.50 (gap) 0.05 head 0.97   hc0 norm 0.47   hc0 down 0.79   hc0 up 0.81 | QSA layers: q8+kv-idx 0.11 k/v+norm-rope 0.23 kv+idx append 0.16 q+q-idx 0.75 scores+topk 0.56 kv-resolve 0.36 attention 0.62 gate 0.05 out-proj 0.37 hc-read1+router 0.97 shared+quant 0.45 waitA 0.16 VRAM hits 1.09 waitB 0.30 PCIe grp 0.11 waitCPU 2.55 copy+combine 0.18 (gap) 0.02   hc0 norm 0.10   hc0 down 0.26   hc0 up 0.27 | total 34.15 ms/window over 69 windows
ab_orig_r0.log strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 1.64 conv 0.19 ab 0.38 z 1.06 rec 0.52 out-proj 1.00 hc-read1+router 2.93 shared+quant 1.33 waitA 2.01 VRAM hits 3.53 waitB 0.55 PCIe grp 0.31 waitCPU 8.34 copy+combine 0.53 (gap) 0.05 head 1.01   hc0 norm 0.49   hc0 down 0.81   hc0 up 0.82 | QSA layers: q8+kv-idx 0.11 k/v+norm-rope 0.24 kv+idx append 0.16 q+q-idx 0.78 scores+topk 0.57 kv-resolve 0.38 attention 0.64 gate 0.05 out-proj 0.38 hc-read1+router 1.00 shared+quant 0.45 waitA 0.17 VRAM hits 1.16 waitB 0.20 PCIe grp 0.11 waitCPU 3.48 copy+combine 0.19 (gap) 0.02   hc0 norm 0.10   hc0 down 0.27   hc0 up 0.27 | total 38.26 ms/window over 72 windows
ab_orig_r1.log strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 1.60 conv 0.18 ab 0.37 z 1.03 rec 0.50 out-proj 0.98 hc-read1+router 2.85 shared+quant 1.30 waitA 1.43 VRAM hits 3.39 waitB 0.62 PCIe grp 0.32 waitCPU 0.24 copy+combine 0.48 (gap) 0.05 head 0.99   hc0 norm 0.48   hc0 down 0.79   hc0 up 0.81 | QSA layers: q8+kv-idx 0.11 k/v+norm-rope 0.23 kv+idx append 0.16 q+q-idx 0.76 scores+topk 0.57 kv-resolve 0.33 attention 0.62 gate 0.05 out-proj 0.37 hc-read1+router 0.97 shared+quant 0.45 waitA 0.15 VRAM hits 1.11 waitB 0.26 PCIe grp 0.11 waitCPU 0.07 copy+combine 0.17 (gap) 0.02   hc0 norm 0.10   hc0 down 0.26   hc0 up 0.27 | total 25.56 ms/window over 69 windows
ab_orig_r2.log strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 1.58 conv 0.19 ab 0.37 z 1.03 rec 0.49 out-proj 0.97 hc-read1+router 2.84 shared+quant 1.30 waitA 1.61 VRAM hits 3.39 waitB 0.57 PCIe grp 0.32 waitCPU 6.50 copy+combine 0.49 (gap) 0.05 head 0.98   hc0 norm 0.48   hc0 down 0.79   hc0 up 0.81 | QSA layers: q8+kv-idx 0.11 k/v+norm-rope 0.23 kv+idx append 0.16 q+q-idx 0.75 scores+topk 0.56 kv-resolve 0.33 attention 0.62 gate 0.05 out-proj 0.37 hc-read1+router 0.97 shared+quant 0.45 waitA 0.17 VRAM hits 1.10 waitB 0.26 PCIe grp 0.11 waitCPU 2.95 copy+combine 0.17 (gap) 0.02   hc0 norm 0.10   hc0 down 0.26   hc0 up 0.27 | total 34.78 ms/window over 68 windows
ab_v70_r0.log strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 1.40 conv 0.18 ab 0.37 z 0.87 rec 0.50 out-proj 1.01 hc-read1+router 2.80 shared+quant 1.34 waitA 1.69 VRAM hits 2.81 waitB 0.48 PCIe grp 0.34 waitCPU 7.81 copy+combine 0.51 (gap) 0.05 head 0.69   hc0 norm 0.48   hc0 down 0.76   hc0 up 0.72 | QSA layers: q8+kv-idx 0.11 k/v+norm-rope 0.23 kv+idx append 0.16 q+q-idx 0.68 scores+topk 0.56 kv-resolve 0.38 attention 0.61 gate 0.05 out-proj 0.39 hc-read1+router 0.96 shared+quant 0.46 waitA 0.17 VRAM hits 0.91 waitB 0.20 PCIe grp 0.12 waitCPU 3.31 copy+combine 0.18 (gap) 0.02   hc0 norm 0.10   hc0 down 0.25   hc0 up 0.24 | total 34.91 ms/window over 66 windows
ab_v70_r1.log strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 1.40 conv 0.18 ab 0.37 z 0.87 rec 0.50 out-proj 1.00 hc-read1+router 2.79 shared+quant 1.34 waitA 1.51 VRAM hits 2.82 waitB 0.48 PCIe grp 0.33 waitCPU 0.27 copy+combine 0.50 (gap) 0.05 head 0.69   hc0 norm 0.47   hc0 down 0.77   hc0 up 0.72 | QSA layers: q8+kv-idx 0.11 k/v+norm-rope 0.24 kv+idx append 0.16 q+q-idx 0.67 scores+topk 0.56 kv-resolve 0.37 attention 0.60 gate 0.05 out-proj 0.39 hc-read1+router 0.97 shared+quant 0.46 waitA 0.17 VRAM hits 0.91 waitB 0.21 PCIe grp 0.12 waitCPU 0.10 copy+combine 0.17 (gap) 0.02   hc0 norm 0.10   hc0 down 0.26   hc0 up 0.24 | total 23.96 ms/window over 66 windows
ab_v70_r2.log strata decode GPU stages (ms/window): GDN layers: q8+qkv/q-idx gemv 1.40 conv 0.18 ab 0.37 z 0.88 rec 0.50 out-proj 1.05 hc-read1+router 2.82 shared+quant 1.35 waitA 1.70 VRAM hits 2.89 waitB 0.41 PCIe grp 0.33 waitCPU 5.60 copy+combine 0.50 (gap) 0.05 head 0.70   hc0 norm 0.48   hc0 down 0.78   hc0 up 0.73 | QSA layers: q8+kv-idx 0.11 k/v+norm-rope 0.23 kv+idx append 0.16 q+q-idx 0.68 scores+topk 0.57 kv-resolve 0.40 attention 0.62 gate 0.05 out-proj 0.39 hc-read1+router 0.97 shared+quant 0.46 waitA 0.17 VRAM hits 0.93 waitB 0.15 PCIe grp 0.11 waitCPU 2.98 copy+combine 0.17 (gap) 0.02   hc0 norm 0.10   hc0 down 0.26   hc0 up 0.24 | total 32.50 ms/window over 71 windows
